Overlap scan and closeout: 2026-09-30¶
Requested September 27–30 across four arXiv tracks. All live pages complete with zero cache fallback. Hits 197/1081/569/976 are broad query pools with overlap, not one day's eligible papers. Filtering to v1 after September 29 leaves 8/127/62/78, still needing full text, deduplication and reproducibility review. Institutional pages timed out during TLS read: arXiv complete, cross-source incomplete, watermark unchanged. No DeepXiv or third-party substitutes.
| Paper | Area / priority | Result |
|---|---|---|
| HELIX | Industrial P0 | TikTok e-commerce full-traffic A/B (§3.5 Tables 4–5); three token streams, one-way reusable cache, MPTF |
| STEPQuant | Foundation P1 | Author code; temporal lifetime × spatial impact bit allocation and dual-axis scale |
| LeapQuant | Foundation P1 | Window quantization, compensation tokens and residual smoothing |
| Chinese-Jev | System One P1 | Chinese candidate marker head and CE/RLCD |
| Dr. OPD | Post-training P1 | Author code; dual-level question token credit and weighted OPD |
| SIPO | Post-training P1 | Author code; dual-context contrastive teacher and token advantage |
| ReMem | Agent P1 | Author code; temporal memory and Multi-Memory GRPO |
| Video-RSI | Agent/RSI P1 | Author code; accuracy–visual cost Pareto gate |
PR #180 already implemented ROFT, LSPD, Harness Learning, MS-GLA and GraphHCA; update ledger/roadmap only. RECAP (2609.37905), Google Research, lacks quantified A/B or accepted full traffic and fails the industrial gate. SelfSearch (2609.37968) is deferred for real search, reward-free trajectory filtering and equal-budget agent comparisons. VACE (2609.37105) is deferred for editable multi-tool agents, execution feedback and fair intergenerational cost. Next overlap at least three days and retry Google/Meta/DeepMind; don't call the historical pool new papers or advance before auditable states.